What users should know about Full Disk Encryption based on LUKS

Transcription

1 What users should know about Full Disk Encryption based on LUKS Andrea VISCONTI Department of Computer Science Università degli Studi di Milano BunnyTN15 December 17, / 34

2 Overview 1 Introduction 2 The key management process 3 Linux Unied Key Setup (LUKS) 4 Analysis of a LUKS implementation [1] 5 Conclusions BunnyTN15 andrea.visconti@unimi.it December 17, / 34

3 Problem Description Introduction Mobile devices, laptops, USB memory usually store large amounts of sensitive information frequently unprotected. If such devices are lost or stolen, the risk of unauthorized disclosure of condential, sensitive, or classied information is very high. Also operating systems store temporary les and/or swap partitions on hard drive and several problems arise when these les contain sensitive data. BunnyTN15 December 17, / 34

4 A possible solution Introduction A possible solution is to encrypt the whole hard disk: Full Disk Encryption (FDE). FDE solutions aim to provide data security, even in the event that an encrypted device is lost or stolen. All information is encrypted/decrypted on the y, automatically and transparently. Without the encryption key, the data stored on the disk remains inaccessible to any users (regular or malicious). BunnyTN15 andrea.visconti@unimi.it December 17, / 34

5 Introduction Main issue and possible solution One of the main issues facing Full Disk Encryption solutions is the password management. Indeed, the master key used to encrypt the whole disk is stored on it. A well-known solution to this problem, is to adopt a two level key hierarchy but sometimes it is not enough (e.g. two level key hierarchy adopted by Android 3-4.3). We analyze the key management process used to derive the encryption key; how the choice of specic hash functions and aggressive power management options may aect the security of a FDE solution. BunnyTN15 andrea.visconti@unimi.it December 17, / 34

6 The key management process The key management process LUKS adopts a two level key hierarchy: A strong master key generated by the system is used to encrypt/decrypt whole hard disk; The master key is unique; Such a key is split, encrypted with a secret user key and stored on the device itself; Each user has their own secret key; Several encrypted master keys are stored on disk, one for each user. A possible solution is to adopt a KDF [2, 3]. The KDF implemented is PBKDF2 [4]. BunnyTN15 andrea.visconti@unimi.it December 17, / 34

7 The key management process The key management process (2) DK = PBKDF 2(p, s, c, dklen) More precisely, the derived key is computed as follows: DK = T 1 T 2... T dklen/hlen Each single block T i is computed as T i = U 1 U 2... U c where U 1 = PRF (p, s i) U 2 = PRF (p, U 1 ). U c = PRF (p, U c 1) BunnyTN15 andrea.visconti@unimi.it December 17, / 34

8 The key management process The key management process (3) In order to prevent building universal dictionaries, PBKDF2 uses a salt. The KDF allows legitimate users to spend a moderate amount of time on key derivation, while inserts CPU-intensive operations on the attacker side. BunnyTN15 andrea.visconti@unimi.it December 17, / 34

9 A LUKS partition Linux Unied Key Setup (LUKS) A LUKS partition contains information about start sector of key material; salt; iteration counts (e.g. c = 588, 761); used cipher (e.g. AES); cipher mode (e.g. XTS); key length (e.g. 256 bits); hash function (e.g. SHA-1); master key checksum;... BunnyTN15 andrea.visconti@unimi.it December 17, / 34

10 A LUKS master key Linux Unied Key Setup (LUKS) When a user key is provided, it unlocks one of the eight key slots and the following algorithm is executed: BunnyTN15 andrea.visconti@unimi.it December 17, / 34

11 Linux Unied Key Setup (LUKS) LUKS: the master key recovery process BunnyTN15 December 17, / 34

12 Our testing activities... Because salt parameter is known and user password may be guessed, we focus on iteration counts and their ability to slow down a brute force attack as much as possible. We try to understand 1 where and how the iteration counts are used; 2 how the choice of specic hash functions may aect the iteration count computation; 3 how unwitting users might signicantly reduce the security of a LUKS implementation by setting aggressive power management options. BunnyTN15 andrea.visconti@unimi.it December 17, / 34

13 Our testing activities... 1) Where and how the iteration counts are used... BunnyTN15 December 17, / 34

14 FIRST: Iteration counts (where and how) BunnyTN15 December 17, / 34

15 FIRST: Iteration counts (where and how) (2) This table shows the average iteration count values involved in the key derivation process. BunnyTN15 December 17, / 34

16 FIRST: Iteration counts (where and how) (3) This table shows the average iteration count values involved in the master key checksum process. BunnyTN15 December 17, / 34

17 FIRST: Iteration counts (where and how) (4) A closer look: we experimentally observed that about 75-80% of the computational eort required to compute a derived key is generated by rst iteration count, while the remaining 20-25% by second one. BunnyTN15 andrea.visconti@unimi.it December 17, / 34

18 Avoiding the master key checksum process BunnyTN15 December 17, / 34

19 Our testing activities... 2) How the choice of specic hash functions may aect the iteration count computation... BunnyTN15 December 17, / 34

20 SECOND: Iteration counts and hash functions The iteration counts are automatically computed by making some run-time tests when the encrypted partition is generated. We collected several partition headers to be sure that such values are not conditioned by external factors, e.g. running programs. More precisely, for each processor (eight) and each hash function (four), we execute 100 runs for a total of = 3200 partition headers collected. The variation across runs is observed to be less than 0.4%. BunnyTN15 andrea.visconti@unimi.it December 17, / 34

21 SECOND: Iteration counts and hash functions (2) As expected, devices with a dierent hardware conguration generate dierent iteration count values. Surprisingly, the choice of a dierent hash function may considerably decrease the iteration count values. BunnyTN15 andrea.visconti@unimi.it December 17, / 34

22 SECOND: Iteration counts and hash functions (3) The iteration counts related to SHA-256/512 are considerably smaller than those of SHA-1. It is easier to attack a FDE solution which makes use of a new hash function (e.g., SHA256 or SHA512) rather than one which uses an old hash function (e.g., SHA-1). This is a counter-intuitive idea! From an user's point of view, SHA-256 and SHA-512 are still considered more secure than SHA-1, therefore a FDE solution based on SHA-2 is expected to be stronger. BunnyTN15 andrea.visconti@unimi.it December 17, / 34

23 SECOND: Iteration counts and hash functions (4) The computational time spent to compute a list of master key candidates does not only depend on the iteration count values. Even the number of ngerprints required to compute a single iteration aects the total execution time! BunnyTN15 andrea.visconti@unimi.it December 17, / 34

24 PBKDF2 Analysis of a LUKS implementation DK = PBKDF 2(p, s, c, dklen) More precisely, the derived key is computed as follows: DK = T 1 T 2... T dklen/hlen Each single block T i is computed as T i = U 1 U 2... U c where U 1 = PRF (p, s i) U 2 = PRF (p, U 1 ) U c. = PRF (p, U c 1) BunnyTN15 andrea.visconti@unimi.it December 17, / 34

25 SECOND: Iteration counts and hash functions... but a SHA-1 ngerprint is only 160 bits in length and cannot be used as derived key, hence a second ngerprint is necessary: DerKey = T 1 T 2 On the other hand, SHA-256 and SHA-512 generate enough bits to compute a derived key: DerKey = T 1 At equal iteration count values, we need to run the HMAC-SHA1 two times. Hence, a FDE solution based on HMAC-SHA1 slow down a brute force attack better than one based on HMAC-SHA256/SHA512. BunnyTN15 andrea.visconti@unimi.it December 17, / 34

26 THIRD: Iteration counts and power management 3) How unwitting users might signicantly reduce the security of a LUKS implementation by setting aggressive power management options... BunnyTN15 andrea.visconti@unimi.it December 17, / 34

27 THIRD: Iteration counts and power management Another important feature that users have to take into account during encryption operations are the power management options. A common way to increase the battery life of devices is to enable aggressive power saving policies. Such policies save power, but they also impact performance by lowering CPU clock speed.... and the iteration count values fall down even further! BunnyTN15 andrea.visconti@unimi.it December 17, / 34

28 THIRD: Iteration counts and power management This means that power saving policies might have an important impact on the iteration count values, hence, on the strength of the FDE solution adopted. BunnyTN15 December 17, / 34

29 Testing: a toy example We implemented a brute-force attack based on a password-list of 250,000 master keys (i.e. a dictionary attack). We run our code on a laptop equipped with an i7 4710MQ processor. No GPUs have been used. We target two LUKS partitions collected using the following conguration options: default iteration count values, AES-256 XTS mode, HMAC-SHA1, laptop plugged in; default iteration count values, AES-256 XTS mode, HMAC-SHA512, laptop unplugged; BunnyTN15 andrea.visconti@unimi.it December 17, / 34

30 Testing: a toy example (2) The second approach (i.e., SHA-512 unplugged) abruptly reduce the timeframe for brute forcing, showing how the simple choice of conguration parameters may aect a FDE solution based on LUKS. BunnyTN15 andrea.visconti@unimi.it December 17, / 34

31 Conclusions Conclusions We identify a number of issues that should be assessed and faced when a full disk encryption is implemented: The iteration count values are used to slow down a brute force attack, therefore, they should not be too small. Experimental results show that sometimes they are. The problem of EXT family le system allows attackers to substitute the master key checksum process by a simple decryption operation. The CPU-intensive operations used to compute a derived key should not be avoided by executing a set of functionally-equivalent instructions. From an user's point of view a FDE solution based on HMAC-SHA256/SHA512, is expected to be much stronger than one based on SHA-1, and be far more resistant to brute-force attacks. Our testing disprove this. BunnyTN15 andrea.visconti@unimi.it December 17, / 34

32 Conclusions (2) Conclusions Power management options should not aect the strength of a FDE solution. Testing results show that aggressive power-saving approaches may have a relevant impact on the iteration count values, hence, on the strength of the solution adopted. Master keys stored on disk are protected with user keys which should have a minimum length requirement in order to prevent a brute force attack. We experimentally observed that a number of distribution such as Debian, Ubuntu, and ArchLinux have no minimum length requirement, while Fedora has (but only eight characters)....the cryptsetup 1.7 release changes defaults for LUKS... BunnyTN15 andrea.visconti@unimi.it December 17, / 34

33 Conclusions Thanks for your attention! BunnyTN15 December 17, / 34

34 Bibliography Conclusions S. Bossi and A. Visconti, What users should know about Full Disk Encryption based on LUKS, in Proceedings of the 14th International Conference on Cryptology and Network Security, CANS15, RSA Laboratories, Pkcs #5 v2.1: Password based cryptography standard, NIST, SP : Recommendation for password-based key derivation, A. Visconti, S. Bossi, H. Ragab, and A. Caló, On the weaknesses of PBKDF2, in Proceedings of the 14th International Conference on Cryptology and Network Security, CANS15, BunnyTN15 December 17, / 34